2026
10,780 articles · 136 days · 10 categories · avg 79 per day
SLIC and Semantic Shapley Value: Measuring Agent Contributions in LLM Systems With 93% Less Compute
Researchers introduce SLIC, a method that scores each agent's contribution in a multi-agent LLM system in a single pass and cuts computational cost by 93.3%.
PEARL: A 4B Model Beats DeepSeek-V3.2 685B at Optimization Modeling
PEARL translates natural-language problems into solver code and fixes it in a loop. Its 4B-parameter version outperformed the 685B-parameter DeepSeek-V3.2.
SmolLM2-1.7B Beats 34B Models at MLIR Code Generation Without Fine-Tuning
Researchers show that SmolLM2-1.7B, with decoding constraints derived from MLIR dialect specs, outperforms 34B-parameter models and runs 8-25 times faster.
BatchDAG: LLM Planner Builds an Operation Graph and Cuts Model Calls by Up to 47x
Thinky Releases Open Inkling 975B-A41B Model Under Apache 2.0 License
Evidence Chain Evaluation: An AI Fact-Checker That Is Allowed to Say "Not Sure" Instead of Issuing False Verdicts
SysAdmin Benchmark Measures AI Power-Seeking: Frontier Models Score 0–5%
Lila Sciences: Science as Data for AI — an Autonomous Lab Backed by $550M
Lila Sciences is building autonomous labs where AI runs experiments through robots, betting that science is the last inexhaustible source of training data for…
AI Cybersecurity Moves to the Front of the Agenda, Latent Space Says
Latent Space's daily AINews newsletter flags a trend: a run of fresh headlines has pushed AI cybersecurity to the top of the industry's 2026 agenda.
Xaira unveils X-Cell, a 4.9B-parameter "virtual cell" for drug discovery
Linus Torvalds on AI in Linux: a useful tool, and dissenters can fork
How Programmers Are Adapting to AI: Reviewing Machine-Written Code Instead of Writing It
Thinking Machines Releases Inkling, an Open 975B-Parameter MoE Model
Are AI Labs Training Models for the "Pelican on a Bicycle" Test? A Look at Pelicanmaxxing
Dylan Castillo ran 48 prompts through 7 models to check whether AI labs train their models specifically for Simon Willison's famous pelican-on-a-bicycle benchmark.
OpenAI Models Breached Hugging Face's Internal Systems in Hours Instead of Weeks
Advanced OpenAI models breached Hugging Face's internal systems in a matter of hours, an attack that would have taken a human weeks, Bloomberg Tech reports, citing…
Chinese AI Startup Psibot Raises Nearly $100 Million at a $1.48 Billion Valuation
Chinese startup Psibot is closing a round of nearly $100 million at a $1.48 billion valuation, becoming the latest AI unicorn as investor interest in the sector…
Ant Group Takes Stake in AI Robotics Startup Lexiang Intelligent
Yuewen Unveils Buddy AI Agents for Writers, Comic Dramas and Copyright Work
FutureTech at WAIC 2026: How the "One Person + a Swarm of AI Agents" Model Is Changing Startups
Shanghai Extends STAR Market's Fifth Listing Standard: Loss-Making AI Companies Can Go Public
Tabula: A Privacy-Preserving Foundation Model for Single-Cell Genomics and Cell Rejuvenation Factor Discovery
Researchers unveil Tabula, a foundation model that trains on data from multiple labs without sharing it and proposes rejuvenation factors for human cells.
Ant Digital Unveils Agentar 2.0 at WAIC 2026, a "Super-Factory" for Business Agents
At WAIC 2026, Ant Digital unveiled Agentar 2.0, a "super-factory" platform with nearly 200 role-level digital expert templates and hundreds of ready-made agent…
Memory Merge DQN: Merging Target Network Weights Instead of Hard Replacement in DQN
AI Decodes the Meaning of Speech From Brain Activity and Transfers to New People Without Retraining
Prefix-GRPO: Training Small Language Model Agents on Trajectory Prefixes
OOD Benchmark Leak: How a Contaminated Class Dragged a Detector's AUROC Down to 0.326
SAE Interpretability Evaluation: The Measurement Pipeline Matters More Than the Model Architecture
An arXiv study finds that comparisons of sparse autoencoder (SAE) interpretability are driven mainly by the evaluation pipeline, not by the architecture of the…
FNO and UNO Neural Operators for Calculating Two-Dimensional Neutron Flux
Researchers compared three surrogate models built on Fourier and U-shaped neural operators to approximate two-dimensional neutron flux in radiation shielding…
SUM: A Server-Side Fix for Forgetting in Federated Continual Learning, Up to 22% Gain
SUM corrects "adaptation vectors" on the server and delivers up to 22% gains over earlier federated continual learning methods, with no extra load on clients.
How to Explain Continual Learning in Neural Networks: Attention Rollout and Grad-CAM on Time Series
Air Quality Arena: A Benchmark for Forecasting Air Pollution with Foundation Models
CruiseBench: A New Benchmark for Evaluating Neural Networks on Aircraft Engine Lifetime Prediction
